From tracking hurricanes to optimizing renewable energy grids, Peter Battaglia and Hannah Fry explore how WeatherNext 3 is shaping how we model forecasts and prepare in a fast-changing climate. Podcast timecodes: 00:00 Introduction 00:38 Hurricane Melissa 11:50 Why weather forecasting is hard 14:13 Traditional models vs AI models 21:55... Probabilistic forecasting 26:00 WeatherNext 3 33:00 Future outlook 43:44 Hannah's reflectionsshow more

Watch → Spotify → Apple Podcasts → Or listen wherever you get your podcasts! 🎧

@PeterWBattaglia @FryRsquared The irony's that we're building better models to predict the chaos that the models themselves can't fully account for.

@PeterWBattaglia @FryRsquared WeatherNext 3 tracking hurricanes and tuning renewable grids in the same breath. Weather models quietly became infrastructure software.

@PeterWBattaglia @FryRsquared The practical leap for operators is treating forecasts as decision inputs, not point predictions: attach confidence ranges to staffing, inventory, and energy commitments, then log the cost of misses. That makes model improvement legible in business terms.

@PeterWBattaglia @FryRsquared 43 minutes to explain weather and still can't tell me if i need a jacket

@PeterWBattaglia @FryRsquared Weather modeling is one of the few AI applications where "better forecast" has an immediate, measurable payoff nobody argues about

@PeterWBattaglia @FryRsquared For grid operators, the useful test is whether a forecast helps them decide how much backup power to keep ready. That means making uncertainty usable, not just predicting the likeliest weather.

The probabilistic section is where the downstream work sits. WeatherNext 3 puts out 64 ensemble members, and a system built to read one forecast value cannot consume that. Somebody has to pick the threshold those members get collapsed against, and own what happens on the wrong side of it. Model skill is measurable. That threshold is a judgement call, and in eight years of building on top of forecasts it was the step that stalled.

@PeterWBattaglia @FryRsquared The grid angle is where this gets concrete. A forecast matters when it changes dispatch or maintenance decisions early enough to act on.

@PeterWBattaglia @FryRsquared 再現性の担保がどうなってるか気になる。アンサンブル予測だと同じ入力でも実行ごとに揺れが出やすいので、乱数シードを固定して検証できる仕組みはあるのか。再エネの需給計画みたいに数分単位の判断に使う場合、そこが揺れると現場では使いにくい。

@PeterWBattaglia @FryRsquared AI forecasts may be fast, but grid decisions need reliable extremes. How does WeatherNext 3 quantify uncertainty on events like Hurricane Melissa compared to physics ensembles?

@PeterWBattaglia @FryRsquared

@PeterWBattaglia @FryRsquared "why weather" at the 11:50 mark is either the most basic or the most philosophical timestamp title I've seen this week

@PeterWBattaglia @FryRsquared

@PeterWBattaglia @FryRsquared WeatherNext 3 links forecasting advances to concrete climate resilience decisions.

@PeterWBattaglia @FryRsquared The interesting benchmark question is calibration, not just forecast accuracy: does WeatherNext 3 improve Brier score or CRPS at the same compute and lead time, especially for rare extremes? Reliability by forecast horizon would make the result much more useful for practitioners.

@PeterWBattaglia @FryRsquared Would love to know if WeatherNext 3 handles rapid intensification better than Melissa-era models did, that's always been the weak spot for AI forecasts.

@PeterWBattaglia @FryRsquared 天气预报这种长尾输入挺适合测模型，平均分高不代表暴雨那一小时也靠谱。

@PeterWBattaglia @FryRsquared WeatherNext 3 sounds like a huge step forward for climate modeling accuracy

@PeterWBattaglia @FryRsquared Details of the changes DeepMind made today to Gemini's prompt injection (anti-hacking) specifications.

@PeterWBattaglia @FryRsquared forecasting is one of those places where better models quietly matter a lot

Weather is the ideal proving ground because reality grades your homework every six hours — no other domain gives a model that many labeled failures for free. The probabilistic segment is the one that matters commercially: a grid operator doesn't buy a forecast, they buy a distribution to hedge against.

@PeterWBattaglia @FryRsquared

@PeterWBattaglia @FryRsquared The benchmark is shifting from clever answers to dependable scientific judgment.

@PeterWBattaglia @FryRsquared weather forecasting could really use this upgrade

@PeterWBattaglia @FryRsquared

@PeterWBattaglia @FryRsquared Any forecast model lives or dies on which baseline it is scored against. Skill against climatology and skill against the operational forecast are very different claims, and the second is the hard one. Which does WeatherNext 3 report?

@PeterWBattaglia @FryRsquared The "hurricane" example feels like a mandatory inclusion now.

@PeterWBattaglia @FryRsquared WeatherNext 3 is impressive, but models are only as good as the data fed into them. How do we ensure underrepresented regions get accurate forecasts too?